↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Software factory/

Zeroshot

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1564 words)
agent-curated 592 fully-ai-generated 695 llm=glm-5.3-flash 602 multi-agent 54 research-note 562 review-loop 1 software-factory 19
Table of Contents

Zeroshot is an MIT-licensed CLI that runs a coding task as an explicit graph: a worker agent implements the change, two independent reviewer agents must both accept, and rejected work returns through a bounded repair loop before anything is delivered.

What it is
#

The Open Engine’s Zeroshot is a native Rust binary installed from npm that also drops one skill into Codex, Claude Code, GitHub Copilot CLI, and Pi, with contributor-maintained OpenCode and Hermes adapters beside the core four in local builds, then runs one of those harness CLIs as both the worker and the reviewers with your existing logins. A GraphSpec file defines control flow and typed state, a RuntimePlan binds each node to a harness, provider, model, and named connections, and the run input is validated before execution opens a durable SQLite ledger. The built-in software-change template routes worker output to parallel acceptance and code review, sends rejected work to a dedicated repair worker with both reviews’ feedback, and re-runs both reviews until a retry cap stops the loop. Delivery is a ladder you opt into, from keep-it-local through –push, –pr, and –ship through PR, CI, and merge, and runs execute locally, on a self-hosted Docker target, or on Zeroshot Platform’s managed capacity (Zeroshot Cloud until the v10.18.0 rebrand).

Status
#

Active and shipping: 1,969 stars since creation on 2025-12-25, per the GitHub API as of 2026-10-13. Pushed 2026-10-31, with the train at v10.24.0 (2026-10-31), npm in sync at 10.24.0, and the Python SDK release train (zeroshot-python tags) riding each release, per the GitHub API and the npm registry as of 2026-10-31. v10.14.0 adds Pi as a fourth harness beside Claude Code, Codex, and GitHub Copilot CLI, and v10.13.0 guards merges with renewable delivery credentials that stay out of agent environments, profiles, and history. v10.16.0 extends that credential line with human-managed machine delegations and expiring API keys for unattended runs (delegation management requires a human login, and machine keys stay out of run inputs and subprocesses), and local PR delivery now rejects a target branch absent from the selected remote before agents start. v10.17.0 (2026-10-24) adds Active Memory for local and Docker runs, enabled by default: deterministic hooks recall concise lessons from costly repository mistakes and recurring graph procedures, and a resumable zeroshot memory init builds that memory from local run history. v10.18.0 (2026-10-24) rebrands Zeroshot Cloud as Zeroshot Platform across the CLI, docs, and SDKs, and its notes state the old $100 Cloud pricing is no longer stated anywhere. v10.19.0 (2026-10-26) adds a guided zeroshot-graph-design interview skill to the canonical installed set and moves the Platform target’s hosted endpoints to app.zeroshot.sh, and v10.19.1 fixes the Python SDK’s PyPI publishing while bumping the bundled harnesses (Codex 0.164.0, Claude Code 2.1.310, Copilot 1.0.103, Pi 1.4.0). v10.20.0 through v10.24.0 (2026-10-30 and 2026-10-31) add host context with Git and checkpoint hardening, caller-declared model context and output limits, adapter isolation with contributed OpenCode and Hermes harnesses (contributors maintain them, and Platform independently admits only the core registry), terminal-tree rendering of template graphs, and lesson reconciliation that extends Active Memory to hosted runs. v10.15.0’s README repositions the tool around user-configured workflows, with the built-in review loop moved below them, and strips the Zeroshot Cloud section, its pricing pointer, and the Cloud docs links, with the maintainers stating plainly that they are not advertising Cloud yet.

Star History Chart

The independent footprint is thin: one 2-point Hacker News thread from June 2026 and a young curated list, so the README’s “starred by engineers at Google, Meta, Shopify” badge is a self-published claim. v8 was a hard interface cutover that replaced a Node.js runtime with the native binary, and the README now routes reproducible bugs and feature requests to GitHub Issues and runs contributions through a CONTRIBUTING.md with a Discord claim process, though contributed adapters stay outside the product-maintained core.

Strengths
#

  • The acceptance rule is architectural: reviewer nodes return comments and never edit, the implementing agent cannot approve its own output, and required checks re-run after every repair.
  • The graph defines sequence, parallel steps, retry caps, and exit conditions before execution, so a run’s behavior is inspectable in advance.
  • RuntimePlan hands credentials to one node at a time, and reviewers do not inherit the worker’s keys.
  • Profiles save a graph plus its per-node model bindings for reuse, and custom topologies are a documented extension path.

Cautions
#

  • Acceptance is LLM judgment, not a deterministic gate: the README states coverage depends on the requirements, reviewers, tests, and environment you provide, a weaker guarantee than Fluent’s final Tester or Ouroboros’ mechanical first stage.
  • Local runs edit the current git worktree with no isolation, so isolation means standing up the Docker target.
  • Thin independent evaluation, and the v8 cutover shows interface stability is not a promise.
  • Zeroshot Platform (Cloud until the v10.18.0 rebrand) has no published prices, the v10.15.0 README stopped advertising it entirely, and the v10.18.0 notes confirm the old $100 price is no longer stated anywhere.

Pricing
#

The core is MIT, free, and self-hosted on your harness logins and keys. Zeroshot Platform, the managed target (Zeroshot Cloud until the v10.18.0 rebrand), covers accounts, organization policy, and queue behavior in its docs, without published prices as of 2026-10-25. The v10.15.0 README (2026-10-19) removed the Cloud section and its pricing pointer, and the v10.18.0 release notes state the old $100 price is no longer stated anywhere, so the managed path is currently unadvertised rather than sold.

Compared to
#

  • Ouroboros hides the grading command and expected result from the worker and gates mechanically first; Zeroshot shows reviewers the code but separates who judges from who wrote.
  • Fluent ends in a deterministic Tester bound to the reviewed commit; Zeroshot ends in independent reviewer agents plus your configured checks.
  • Fabro makes you author any DOT graph with human gates; Zeroshot ships the review loop built in and treats custom topology as the extension path.

Bottom line
#

Recommended for engineers who want issue-to-PR runs where review is structurally independent from implementation, on harnesses they already pay for. Not for teams that require a deterministic final gate, third-party validation first, or interface stability guarantees.

Changes
#

  • 2026-10-10 - Created from the awesome-software-factories scan as the independent-review-loop member.
  • 2026-10-16 - Recorded the train to v10.12.0 (2026-10-15, npm in sync), the new Python SDK release train, and the two features (planner-led archived exploration with three independent scouts in v10.11.0, local-folder and generic-Git-transport runs in v10.12.0).
  • 2026-10-21 - Recorded the train to v10.15.0 (2026-10-19, npm in sync): Pi added as a fourth harness in v10.14.0, renewable delivery credentials kept out of agent environments in v10.13.0, and the v10.15.0 README repositioning that leads with user-configured workflows and stops advertising Zeroshot Cloud; the What-it-is harness roster and the Pricing and caution lines updated.
  • 2026-10-23 - Recorded the train to v10.16.0 (2026-10-22, npm in sync): machine delegations with expiring API keys for unattended runs, human-gated delegation management, keys kept out of run inputs and subprocesses, and a remote-base preflight before local PR delivery.
  • 2026-10-25 - Recorded the train to v10.18.0 (2026-10-24, npm in sync): Active Memory in v10.17.0 (lessons recalled from costly repository mistakes on local and Docker runs, enabled by default) and the Zeroshot Cloud to Zeroshot Platform rebrand in v10.18.0, whose notes state the old $100 Cloud pricing is no longer stated anywhere; the What-it-is run-target line and the Pricing and caution lines updated.
  • 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
  • 2026-10-26 - Corrected the star-history chart’s light source, which wrongly carried the dark-theme parameter.
  • 2026-10-28 - Recorded the train to v10.19.1 (2026-10-26, npm in sync): the guided zeroshot-graph-design interview skill and the Platform endpoint move to app.zeroshot.sh in v10.19.0, and Python SDK publishing fixes plus bundled-harness updates in v10.19.1.
  • 2026-10-31 - Recorded the train to v10.24.0 (2026-10-31, npm in sync): host context with Git and checkpoint hardening (v10.20.0), caller-declared model context and output limits (v10.21.0), adapter isolation with contributed OpenCode and Hermes harnesses (v10.22.0, contributors maintain them, Platform admits core only), template-graph terminal trees (v10.23.0), and Active Memory lesson reconciliation extended to hosted runs (v10.24.0); the README now links GitHub Issues and a CONTRIBUTING.md, replacing the declining-contributions claim, and the What-it-is harness roster updated.

See also
#

  • Ouroboros - the other anti-self-grading member, hidden grading versus independent reviewers
  • Fluent - the deterministic final gate comparison
  • Fabro - the graph-authoring alternative
  • Software Factory Feature Matrix - where the review-loop column sits among the factories

References
#